gpu-unlock removes a GeForce speed cap for AI, and has 17,239 stars

NVIDIA sells consumer GeForce cards with their tensor cores capped at half speed. The cap applies to the math that nearly all AI training uses. A Linux kernel module called gpu-unlock removes the cap at boot. It runs before the stock NVIDIA driver loads. The project is on GitHub as Zagitta/gpu-unlock. The GitHub API returned 17,239 stars for it on 6 November 2026.

The repository was created on 18 October 2026, per the GitHub API. The project is GPL-2.0-only. It has 1,736 forks and 121 open issues.

What the cap is, and what the fix does

The README says GeForce RTX 20, 30 and 40 cards are capped at half rate. The cap applies to FP16, BF16 and TF32 math with FP32 accumulation. Those are the formats most AI frameworks use. The same silicon runs at full rate on NVIDIA's RTX A-series and professional cards, per the README.

gpu-unlock writes two speed selectors to the card at boot. It then does a Function Level Reset. All of this happens before the NVIDIA driver binds the card. The README's own line is that it is "neither a UEFI application nor a patched NVIDIA driver". The author says the module needs no driver patch and no VBIOS flash. A new NVIDIA driver release does not undo it.

The numbers Zagitta reports

The README publishes a table of dense BF16 matrix multiply speed. The numbers are measured with FP32 accumulation at stock power limits. An RTX 3060 12 GB runs at 27 TFLOPS before and 51 TFLOPS after, an 87 percent gain. An RTX 3080 10 GB runs at 65 TFLOPS before and 113 TFLOPS after, a 73 percent gain. An RTX 3090 Ti at 450 W runs at 88 TFLOPS before and 157 TFLOPS after. The RTX 4090 doubles its peak MMA rate, from 165 to 329 TFLOPS. The README notes that the 4090 hits its power limit on large multiplies and sees less in practice.

For real work, the README gives ranges from user reports. LLM prefill in FP16, BF16 or AWQ or GPTQ form sees 25 to 55 percent faster processing, Zagitta says. Time to first token drops 25 to 30 percent. LoRA and full fine-tuning in BF16 see 25 to 50 percent faster passes. Image generation in FP16 or BF16 sees 30 to 55 percent. FlashAttention and SDPA kernels see 50 to 60 percent.

Some workloads see no change, by Zagitta's own account. Token generation after the prompt is memory bound and does not benefit. INT8, FP8 and INT4 quantized models do not benefit. Gaming and plain FP32 CUDA do not benefit. The README says output values are bit-identical before and after; only speed changes.

Costs and caveats

The project is Linux only. The Blackwell RTX 50 generation is not supported, per the README. Laptop cards only stay unlocked if the dedicated GPU is never fully powered off. There is no official NVIDIA endorsement. This is an unofficial driver-adjacent kernel module from a single author. A team running AI work on consumer GeForce cards should decide whether that risk is acceptable first.

Source

GitHub, Zagitta/gpu-unlock README.